AI is gaining traction across the insurance industry, with businesses exploring its potential to reshape everything from underwriting and pricing to claims management. However, turning that interest into practical applications is proving challenging, as insurers grapple with the technology, data and infrastructure needed to make it work.
Fragmented data, ageing systems and disconnected workflows are among the issues standing in the way. There are also questions around cybersecurity, accountability and how much of the existing insurance infrastructure needs to change before AI can be used effectively.
FinTech Global spoke to Melanie Hayes, co-founder of KYND, Matt Twist, vice president of sales, EMEA, at Earnix, and Dan Simmons, CEO of Quensus, to understand what is slowing adoption and what insurers need to address before taking the technology further.
Getting AI into everyday operations
There is no single answer to how insurers should introduce AI. Some may choose to test new applications alongside existing systems, while others will look to incorporate them directly into their core technology. Much depends on the organisation, its existing infrastructure and the decisions it wants the technology to support.
For Hayes, however, the rush to adopt AI could mean insurers overlook some of the more basic security questions.
She explained, “What we do see, looking at a great many organisations from the outside, is that everyone is racing to get ahead of AI adoption, or at least to catch up with it. In that rush, the basic cyber hygiene that should sit underneath it is easy to miss.”
The figures offer some indication of the problem. IBM’s 2026 Cost of a Data Breach Report found that unsanctioned AI tools were involved in 43% of breaches, up from 20% a year earlier. It also found that 68% of breached organisations had no AI governance policy in place.
Hayes believes insurers need to understand where AI is being used, what information it can access and who is responsible for overseeing it. This is not just an internal issue. Businesses applying for insurance may also be using AI without properly understanding the risks, leaving insurers to assess exposures they might not know exist.
As Hayes put it, “AI use that nobody has identified is exposure that nobody has priced.”
Twist agrees that simply adding applications is unlikely to deliver much change if they remain separate from the processes employees use every day. He believes insurers need to integrate AI into the workflows where decisions are actually made, rather than expecting staff to move information between systems and manually check the technology’s recommendations.
He said, “A standalone tool can demonstrate potential, but its value is limited if employees must manually transfer information and reconcile its recommendations with established processes.”
For insurers, this means deciding what they want AI to achieve and establishing who is responsible for the results. A successful trial does not necessarily mean a system is ready to be used across an organisation, particularly when it comes to decisions involving pricing, underwriting or customer claims.
Simmons takes a similar position, arguing that AI needs to be connected to the systems responsible for running an insurer’s business. He points to usage-based underwriting and parametric claims settlement as areas where the technology could have a greater role, but says this will depend on how closely it is integrated with underwriting and policy administration.
Why legacy systems are holding insurers back
Getting AI into everyday workflows is made more difficult by the way insurance businesses have traditionally managed their data and technology.
A single underwriting decision can involve several departments, applications and information sources. These systems may have been introduced years apart and designed for different purposes, leaving insurers with a complicated picture of how information is collected, shared and used.
Hayes said, “A single underwriting decision can pass through several systems, data sources, and functions, each built at a different time for a different purpose.”
The scale of the problem is reflected in AutoRek’s 2026 Insurance Report, which found that insurers manage an average of 17 data sources in their premium processes alone. Meanwhile, 39% identified fragmented data environments as a barrier to AI adoption.
Introducing new technology does not necessarily resolve these problems. If information is incomplete or spread across disconnected systems, AI can inherit the same issues rather than fix them. There is also the question of institutional knowledge. Employees may understand how a particular process works, including the exceptions and decisions that are not formally recorded anywhere.
Hayes explained, “It can only automate a workflow that someone can describe clearly, and that description often lives in people’s heads rather than in any system.”
Security is another consideration. Older systems can be more difficult to patch and monitor, while connecting new applications can create additional routes into sensitive customer information. Hayes also points to the difficulty of making technical data useful to insurance professionals. Information collected for IT security teams may not give an underwriter a clear understanding of the potential financial loss.
She said, “Security data built for IT teams tells an underwriter what exists, but not what matters to a loss.”
Twist believes the challenge extends beyond the technology itself. Integrating AI can expose unresolved questions about data quality, decision-making authority and responsibility across different teams. He argues that insurers need to establish a clear business objective before beginning, while involving the employees who understand the processes being changed.
This is particularly important when deciding which tasks can be automated and which still require approval. Without clear ownership, it becomes difficult to establish who is responsible when a system makes an incorrect recommendation or a process does not work as intended.
Simmons highlights a different but related problem. Many insurance systems were designed around information collected at annual renewals, rather than the continuous streams of data now available through connected devices and application programming interfaces (APIs).
He explained, “The primary operational challenge is that legacy IT systems were originally architected to handle static data at annual renewals, not to ingest continuous, real-time streams of IoT telemetry via APIs.”
For insurers looking to use Internet of Things (IoT) technology to monitor properties and other insured assets, this creates a significant limitation. Even if a device can collect information in real time, the systems responsible for underwriting and managing a policy may not be able to receive or act on it.
Does insurance need a technology overhaul?
The extent of the infrastructure changes needed to support AI is another point of disagreement.
Twist believes insurers can make progress without replacing their entire technology estate. Instead, he suggests focusing on the specific barriers preventing information from reaching the right people and systems.
He said, “Insurers can achieve meaningful value without wholesale infrastructure replacement. However, capturing AI’s full value requires reliable access to relevant data and the ability to feed decisions back into operational workflows.”
A phased approach could allow insurers to improve individual processes while maintaining the systems they already rely on. The important point, according to Twist, is that each change should deliver a practical improvement, whether that means better decision-making, faster processes or improved customer outcomes.
Simmons is less convinced that incremental changes will be enough. He argues that insurers will need to make more substantial investments if they want to use AI to predict and prevent losses rather than simply respond after something has happened.
He said, “The ultimate value of AI lies in moving the industry from a ‘detect and repair’ model to a ‘predict and prevent’ model. This requires a real-time data ecosystem.”
For example, connected devices could provide insurers with information about changing conditions within an insured property. But to make use of that information, insurers need systems capable of receiving it continuously and responding when circumstances change. Simmons believes this requires modern infrastructure, open data protocols and continuous API integrations.
Hayes’ comments highlight why security needs to remain part of those decisions. Connecting more systems and sharing more information may make it easier to use AI, but insurers also need to understand where that information is going and how it is being protected.
The amount of change required will ultimately depend on what an insurer wants to achieve. Improving a back-office process may only require targeted integration, while using connected devices to monitor risks and prevent losses will demand a much greater level of connectivity.
What would an AI-native insurer look like?
While the contributors differ on how much infrastructure needs to change, they agree that insurers cannot treat AI as a separate tool and expect it to transform their operations.
For Hayes, the distinction starts with whether technology has been designed for insurance in the first place. General-purpose applications may need considerable adaptation before their information can be used in underwriting, particularly where accuracy and explainability are important.
She said, “AI-native should mean the same thing. It starts from the decision: what the underwriter needs to see, in what form, and how it can be explained to a reinsurer or a regulator afterwards.”
Twist describes a similar need for AI to become part of the wider decision-making process. Information from pricing, underwriting and customer engagement should be available across the business, rather than remaining within separate systems.
However, he stresses that greater automation must not come at the expense of accountability. Insurers need to know who is responsible for automated decisions, what limits apply and when a person should step in. Some processes may be suitable for greater automation, while others will continue to require human judgement.
Simmons offers a more fundamental picture of how the industry could change. He believes insurers could move away from acting primarily as providers of financial compensation after an incident and towards taking a more active role in preventing damage.
An insurer covering a commercial property, for example, could use real-time information from connected devices to identify signs of a water leak. A smart shut-off valve could then isolate the water supply before the damage becomes more serious. For straightforward claims, smart contracts could also be used to trigger payments when agreed conditions are met.
Simmons said, “An AI-native operation functions as a continuous resilience manager rather than a passive indemnity provider.”
Whether that model becomes widespread will depend on the reliability of connected devices, the quality of the information they provide and insurers’ ability to respond when a risk is identified. For now, the biggest obstacles are less about finding new uses for AI and more about addressing the systems and processes insurers already have. Fragmented data, legacy infrastructure and unclear responsibility all make it harder to move beyond individual trials.
Insurers may not need to replace everything they have built, but they will need to understand where existing systems are falling short. Without that groundwork, introducing more AI is unlikely to solve the problems holding the industry back.
Copyright © 2026 FinTech Global
